Does the teacher-forcing failure generalize to standard text-generation tasks?

Determine whether the training-time failure mechanisms of next-token prediction under teacher-forcing—specifically, Clever Hans cheating and the resulting indecipherable early tokens—observed on the path-star graph path-finding task, generalize to run-of-the-mill text-generation tasks.

Background

The paper introduces and empirically demonstrates a training-time failure of next-token prediction under teacher-forcing on a minimal lookahead task (path-finding on path-star graphs). The authors identify two mechanisms: Clever Hans cheating (where later tokens are trivially fit using previously revealed ground-truth tokens) and the indecipherable token problem (early tokens lose crucial supervision and become hard to learn). These failures cause in-distribution breakdowns for both Transformer and Mamba architectures.

While the study focuses on a controlled planning-style task, the broader significance hinges on whether similar failures occur in common text-generation scenarios that may not explicitly require lookahead or planning. Establishing generalization to everyday text-generation would have important implications for the robustness of teacher-forced training paradigms.